Seven X signals: Jalapeño results, signed-in agents, and a $100 Business seat

Chips and inference hardware

1. OpenAI publishes Jalapeño's first measured results

• What happened: On August 25, OpenAI reported first measured results for Jalapeño, its custom inference chip. On the public InferenceX benchmark, OpenAI says Jalapeño delivered about 1.5–1.9× more AI work per watt at peak throughput and about 1.7–3.6× lower end-to-end latency than the comparison systems across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. The company plans to begin deploying Jalapeño inside its own compute infrastructure by year-end, with Gen 2 already in development. 12
• Why it matters: OpenAI is claiming both higher throughput per watt and lower latency in one architecture, and it is tying those gains to faster ChatGPT responses, more responsive Codex sessions, and agent workloads that compound delay across many steps. 23
• Signal: The numbers come from OpenAI's own tests against commercial systems on InferenceX, with package power ratings used for normalization. Jalapeño is still heading into production qualification and software maturity before the planned year-end deployment. 2

2. Greg Brockman calls AI-assisted chip design underrated

• What happened: On August 26, Greg Brockman wrote that AI for chip design is underrated, quoting a post that OpenAI kept the chip team small and used AI in an improvement loop. OpenAI's Jalapeño write-up says AI helped move the design from initial work to tapeout in nine months, and that AI-generated kernels for selected GPT-OSS blocks ran 1.5–1.8× faster than some human-expert implementations. 245
• Why it matters: The claim is not only that Jalapeño is fast. It is that model-assisted design and programming shortened the loop that produces the next accelerator.
• Signal: Brockman's line is a short endorsement. The nine-month timeline and kernel speedups are OpenAI's own engineering claims. 24

Agents that can act in real accounts

3. ChatGPT Work can sign into websites without seeing passwords

• What happened: Greg Brockman highlighted a ChatGPT product update: ChatGPT Work can now use its computer and browser to sign into websites on web and mobile without ChatGPT seeing the username or password. The product post lists tasks such as booking a DMV or passport appointment, checking insurance reimbursements, submitting invoices to accounting software, and filling small-business permit applications. 67
• Why it matters: Secure login is the missing step between an agent that drafts text and an agent that finishes paperwork on sites that require an account.
• Signal: The announcement is a product claim about the sign-in path and example tasks. It does not report completion rates, failure modes, or which sites are supported. 67

4. Scheduled tasks can fire when Slack, Gmail, or GitHub changes

• What happened: Brockman also flagged triggered tasks in ChatGPT. Plus and Pro users can set ChatGPT Work tasks to run when something changes in Slack, Gmail, or GitHub, not only on a fixed schedule. Scheduled tasks are also rolling out to Free users, who can create up to three tasks, and users can share task setups for others to customize. 89
• Why it matters: Event-driven tasks turn agents from calendar reminders into watchers on the tools where work already arrives.
• Signal: The post describes product availability and caps. It does not publish reliability metrics for event triggers. 89

Business packaging and open tooling

5. OpenAI adds a $100 ChatGPT Business Premium seat

• What happened: OpenAI introduced ChatGPT Business Premium Seats at $100, aimed at small businesses and startups. Greg Brockman summarized the practical difference as no 5-hour limit and more usage. OpenAI's pricing page describes the Premium seat as offering 5× more usage than Standard, with no 5-hour limit, while Business plans remain for teams of 2–200. 101112
• Why it matters: Higher usage ceilings matter for teams that hit rate limits during agent and coding sessions, without moving the whole workspace to Enterprise sales.
• Signal: OpenAI's post is promotional. Exact monthly versus annual dollar amounts on the pricing page did not render cleanly in the captured page text, so treat $100 and the no 5-hour limit / 5× usage claims as the verified product signal. 101112

6. Andrew Ng ships OpenWorker security coworkers

• What happened: Andrew Ng announced a new OpenWorker release focused on security workflows. OpenWorker is an open-source desktop agent that completes tasks on a laptop rather than only chatting. The new build adds built-in cybersecurity agents for scanning code vulnerabilities, checking dependencies for supply-chain injections, and reviewing cloud security configuration. Ng stresses that the harness is open source so teams can audit it, and that users can run open-weight models fully locally so sensitive code stays on the machine. 131415
• Why it matters: The release frames defender tooling as a first-class open-source agent job, with local models as the path for work that closed models may refuse.
• Signal: Ng is describing product features and intent. The post and site do not publish independent evaluation numbers for the security agents. OpenWorker remains labeled beta. 1314

How fast agent product models age

7. Ethan Mollick notes that Agent Builder is already gone

• What happened: Ethan Mollick pointed to OpenAI's October 2025 vision for enterprise AI agents and wrote that OpenAI killed Agent Builder in June. He added that a surprising number of enterprise products still use that older agent model. 16
• Why it matters: Product packaging for enterprise agents is moving faster than many internal builds. A stack designed around last year's builder may already be a legacy pattern.
• Signal: Mollick is offering a practitioner observation about product turnover. The post does not inventory which enterprise products still use the older model. 16